Andrew Fribush
fig. 2.2.9 · running totals query by query, and the query from which screening is ahead.
ARXIVCOCO0.8 → Q150.9 → Q90.925 → Q80.95 → Q80.85 → Q240.9 → Q140.95 → Q11ONE COLUMN IS ONE QUERY: ○ DENSE RUNNING TOTAL, ● SCREENING RUNNING TOTAL, THE RUNGBETWEEN THEM THE GAP, DASHED WHILE SCREENING IS BEHIND05001,0001,500MSSCREENING STARTS 288 MS BEHINDAND GAINS 38.2 MS A QUERYDENSESCREENING151520259QUERIES · DRAG TO READ ANYAFTER 9 QUERIESDENSE 693 MSSCREENING 675 MSSCREENING AHEAD BY 18 MS

Screening here is the route the selector chose, with its calibration and preparation counted in the first query. 0.925 and 0.85 are the predeclared holdout cells.

First use is timed after route warmup and excludes cold context and JIT startup, so the real crossing is later than the one drawn, not earlier.

source: gpu-threshold-join portfolio-selector-2026-09-18 existing/summary.md, holdout/summary.md (RTX 4090, 7d46ea5) · breakeven.data.ts

View in context: Knuth's Root of All Evil Has a Break-Even Point → · All figures

Andrew Fribush, software and computer engineer in Austin. I build consequential systems where software meets institutions: finance, healthcare, commerce, government data, and the mechanisms that govern AI agents. Taking new engagements: hello@andrewfribush.com.